Featured incident
The $500M Claude Bill: A Postmortem
A reported $500M single-month Claude bill — what (reportedly) happened, why agent token spend explodes, and the spend cap that would have stopped it.
writing from the team building Arc
Notes on keeping agents from spending — or breaking — what they shouldn't.
Postmortems, patterns, and the occasional argument about where a guardrail actually belongs. We meter agents and actions, not seats — and we write about why.
Start guarding actions
Sign up free, no credit card. Read the docs, integrate when it earns it.
More posts
- 01 spend-caps
Usage dashboards vs. enforced spend caps
Dashboards report spend after it happens. An enforced cap needs a reserve-then-commit ledger — and naive balance checks race under concurrent agents.
5 min read Read →
- 02 permissions
allow / ask / block: a permission model for agents
Why AI agents need default-deny permissions with a third verb — ask — that firewalls and IAM never had, and what that looks like in code.
5 min read Read →
- 03 human-in-the-loop
Human-in-the-loop: when should an agent ask first?
A decision framework for agent approvals — irreversibility × blast radius × cost — and how it maps to allow / ask / block policy in code.
5 min read Read →
- 04 spend-caps
Why Microsoft cut internal Claude Code licenses
What Microsoft's May 2026 Claude Code cutoff and Uber's blown AI budget say about ungoverned agent spend — and where a cap actually has to live.
5 min read Read →
- 05 agent-delegation
Your agent doesn't need a login. It needs delegated authority.
Sessions, API keys, and OAuth scopes all break for AI agents. What breaks, why, and the four properties delegated authority requires instead.
6 min read Read →
- 06 spend-caps
How to set a spend cap on AI agents — before the bill
Three ways to cap AI agent spend (gateway budgets, provider limits, action-layer caps), what each one actually stops, and working code for an enforced cumulative cap.
3 min read Read →
- 07 agentic-sso
Agentic SSO: what Auth 3.0 means when the user is an AI agent
What agentic SSO and Auth 3.0 mean: why 'who is the human' isn't enough for AI agents, and the 9 questions an AI-native app has to answer.
7 min read Read →
- 08 guardrails
AI Agent Guardrails: The Complete 2026 Guide
The complete 2026 guide to AI agent guardrails: allow/ask/block policy, spend caps, human approval, signed execution, and audit — with patterns and code.
9 min read Read →